Treffer: SEMANTIC SEGMENTATION FOR AERIAL IMAGES

Semantic segmentation of aerial imagery plays a critical role in modern urban planning, environmental monitoring, and the development of smart cities. This project presents an interactive webbased application that performs semantic segmentation on high-resolution aerial images using a deep learning-based U-Net model. The system is developed using Python and integrated into a Streamlit framework to provide a seamless user experience through a browser-based interface. The application allows users to upload aerial or satellite images and visualizes pixel-wise segmentation results across six predefined classes: Buildings, Roads, Land, Vegetation, Water, and Unlabeled. It goes beyond basic segmentation by offering advanced interactive features such as zooming, class mask toggling, and real-time class-wise statistical analysis, including area coverage. The model is trained and evaluated using the “Semantic Segmentation of Aerial Imagery – Dubai, UAE” dataset, which contains pixel-annotated satellite imagery. The proposed system addresses limitations in existing solutions, such as lack of interactivity, low accuracy, and absence of class-wise analytics and toggling class maks. By streamlining the segmentation workflow and offering rich visualization and analytical tools, the system enhances accessibility for non-technical users and supports data-driven decision-making in geospatial analysis.